Discussion
We found that the associations between sleep duration and death were different across levels of physical activity. Short sleep duration was associated with increased mortality risk and shorter survival throughout all categories of daily physical activity. Long sleep duration was associated with higher mortality risk and shorter survival among only participants with a low level of physical activity. Results were consistent when limiting the stratification to leisure time physical activity (walking/bicycling) or work-related activities (work/occupation).
In recent years, several studies have examined the association between sleep duration and death from all causes. A systematic review by Cappuccio et al. concluded that the association is well described by a U shape, with the lowest mortality rates corresponding to 7 hours of sleep per day, and concluded that both short and long durations of sleep are significant predictors of death. The association between short duration of sleep and death is widely accepted, and potential biological mechanisms to explain the effect have already been proposed. There is no agreement, however, on the association between long sleep duration and death, and no studies have proposed any biological mechanism to explain the association. It has been suggested that this association might be a consequence of health disorders that influence both sleep duration and death, and that the association may be hence explained by residual confounding and comorbidities. Stranges et al. evaluated a set of sociodemographic and lifestyle factors that could confound or mediate the U-shaped association between sleep duration and death. They found a significant association between long sleep duration and low physical activity level, which is known to be a strong predictor of death. There are 2 studies that have tried to explore the association between long sleep duration and death according to physical function and health status, both concluding that the association does not substantially change across the considered strata. Mesas et al. evaluated a relatively small cohort of elderly Spanish participants and focused the stratification on health status indicators. Kakizaki et al. analyzed a larger cohort of Japanese men and women and stratified data by physical function coded as limited/unlimited. However, none of these studies examined the association between sleep duration and death according to daily level of physical activity. To the best of our knowledge, the present study is the first to examine the association between sleep duration and death across categories of total physical activity. We documented a lack of association between long sleep duration and death among physically active participants. Our findings seem to support the hypothesis that the association between long sleep duration and death might be partly explained by comorbidity with low physical activity.
When interpreting our results, it is useful to translate tertiles of total physical activity into real-life activities. On average, participants in the lowest tertile hardly ever walk or bike, exercise less than 1 hour/week, spend more than 3 hours/day in inactive leisure time activities (watching television/reading), do less than 1 hour/day of housework, and work mostly sitting. Participants in the highest tertile were, on average, men and women who walked/biked more than 40 minutes/day, exercised more than 3 hours/week, read or watched television less than 1–2 hours/day, did more than 1 hour/day of housework, and had nonsedentary jobs.
Important strengths of this study are the population-based and prospective design and the large sample size, which are features that can increase the generalizability of the results. Moreover, cases were ascertained through direct links with Swedish National Register of Death Causes. It has been estimated that 93% of all deaths in Sweden are reported within 10 days, and 100% are reported within 30 days.
We presented our findings both in terms of mortality rates and in terms of differences in survival time. Reporting results in terms of event rate is the most popular approach in epidemiologic studies, even if these measures are known to be difficult to interpret at the individual level. The percentile of survival links the probability of experiencing the event of interest to a specific time point and represents an important complement to the information provided by risks or rates. Modeling percentiles of survival provides additional information, such as the probability of experiencing an event after a given time period, which may make results easier to interpret and facilitate conveying them to the general public. Different methods to estimate survival percentiles in observational studies are available, but have some limitations that limit their practical application. The well-known Kaplan-Meier estimator is the most commonly used method to derive the survival curve. However, this nonparametric approach is applicable only at the univariate level and does not allow adjustment for potential confounders. In the multivariable situation, survival percentiles can be indirectly derived from semiparametric (e.g., Cox) or parametric methods (e.g., flexible parametric models). These postestimation calculations, however, are not straightforward, especially for deriving confidence intervals for survival percentiles.
Laplace regression, the method we used for statistical analysis, overcomes many of these limitations and provides an easy and flexible tool for estimating percentiles of survival. This method focuses on the time between entry into the study and experiencing the event of interest and directly estimates differences in survival percentiles according to levels of the exposure, providing an intuitive measure of the observed association between the exposure of interest and death directly in the unit of the time scale (i.e., months, years). Laplace regression allows the inclusion of multiple covariates and interaction terms, modeling multiple percentiles simultaneously, and testing regression coefficients within and between survival percentiles. Differences in survival percentiles, as well as hazard ratios, should be interpreted in the context of follow-up length. When the hazard functions are not perfectly proportional, relative measures such as hazard ratios depend on follow-up length. On the contrary, differences in survival percentiles are not influenced by observational period. The 15th percentile is the time by which the first 15% of the population died. Longer follow-up would result in a higher proportion of deaths and the possibility of estimating higher percentiles, but the 15th percentile would not change.
Our study indicates that physical activity level modifies the association between long sleep duration and death. Participants in our cohort had similar characteristics across physical activity levels, limiting the possibility of some source of residual confounding. However, the study has limitations that could have affected the results. The major limitation of our study is the lack of specific questions about sleep (e.g., sleep apnea disorders, quality of sleep, daytime sleep, nighttime awakening). Sleep apnea, in particular, has been related to overall death and might partly explain the association between long sleep duration and death. However, our analyses were adjusted for body mass index and age, which could have accounted for most, if not all, influences of sleep apnea. Another factor that was not included in the questionnaire was depression, which may result in both short and long sleep duration and has been proposed as a latent disease that may partly explain the association between long sleep duration and death. We also lacked information on employment status, which can be related to both sleep duration and physical activity level. Another limitation of the study was that information on sleep duration was self-reported, leading to a potential misclassification of the exposure. In our prospective study, however, classification errors are expected to be nondifferential with respect to the occurrence of death and would most likely attenuate the results.
In conclusion, we found that the association between sleep duration and death was different across physical activity levels. Short sleep was associated with increased mortality risk and shorter survival in all categories of physical activity. Long sleep duration was associated with higher mortality risk and shorter survival among only those participants with low physical activity.